Added value of regional climate modeling over ... - Wiley Online Library

PUBLICATIONS Journal of Geophysical Research: Atmospheres RESEARCH ARTICLE 10.1002/2014JD022781 Key Points: • RCMs produce substantial added value over complex topography regions • High resolution improves precipitation spatial patterns simulation and extremes • RCMs are important tools for climate studies over complex topographical regions

Supporting Information: • Figures S1–S8 and Table S1 Correspondence to: Cs. Torma, [email protected]

Citation: Torma, Cs., F. Giorgi, and E. Coppola (2015), Added value of regional climate modeling over areas characterized by complex terrain—Precipitation over the Alps, J. Geophys. Res. Atmos., 120, 3957–3972, doi:10.1002/ 2014JD022781. Received 30 OCT 2014 Accepted 4 APR 2015 Accepted article online 9 APR 2015 Published online 12 MAY 2015

Added value of regional climate modeling over areas characterized by complex terrain—Precipitation over the Alps Csaba Torma1, Filippo Giorgi1, and Erika Coppola1 1

Earth System Physics Section, The Abdus Salam International Centre for Theoretical Physics, Trieste, Italy

Abstract We present an analysis of the added value (AV) of downscaling via regional climate model (RCM) nesting with respect to the driving global climate models (GCMs). We analyze ensembles of driving GCM and nested RCM (two resolutions, 0.44° and 0.11°) simulations for the late 20th and late 21st centuries from the CMIP5, EURO-CORDEX, and MED-CORDEX experiments, with a focus on the Alpine region. Different metrics of AV are investigated, measuring aspects of precipitation where substantial AV can be expected in mountainous terrains: spatial pattern of mean precipitation, daily precipitation intensity distribution, and daily precipitation extremes tails. Comparison with a high-quality, fine-scale (5 km) gridded observational data set shows substantial AV of RCM downscaling for all metrics selected, and results are mostly improved compared to the driving GCMs also when the RCM fields are upscaled at the scale of the GCM resolution. We also find consistent improvements in the high-resolution (0.11°) versus medium-resolution (0.44°) RCM simulations. Finally, we find that the RCM downscaling substantially modulates the GCM-produced precipitation change signal in future climate projections, particularly in terms of fine-scale spatial pattern associated with the complex topography of the region. Our results thus point to the important role that high-resolution nested RCMs can play in the study of climate change over areas characterized by complex topographical features.

1. Introduction The issue of added value (AV) is central to the use of regional climate modeling as a dynamical downscaling method for the provision of fine-scale regional climate information [e.g., Leung et al., 2003; Giorgi, 2006; Laprise et al., 2008; Laprise, 2014]. Nested Regional Climate Model (RCM) simulations are indeed useful as climate downscaling tools to the extent that they add valuable information to that produced by Global Climate Models (GCMs). The evaluation of the AV of RCM simulations is however not easy, as it often depends on the scale, region of interest, variable, specific application, and experiment configuration. In addition, the assessment of the AV often requires the use of high-resolution, high-quality observational data sets, which are not available for many regions of the globe. The issue of AV has been addressed by a number of authors in different contexts [e.g., Giorgi et al., 1994; Feser, 2006; Castro et al., 2005; Sotillo et al., 2005; Kanamitsu and Kanamaru, 2007; Laprise et al., 2008; Winterfeldt and Weisse, 2009; De Sales and Xue, 2010; Prömmel et al., 2010; Coppola et al., 2010; Racherla et al., 2012; Di Luca et al., 2012, 2013; Diaconescu and Laprise, 2013; Mariotti et al., 2011, 2014; Laprise, 2014; Kerkhoff et al., 2014]; however, due to the complex and multifaceted aspects of the problem, general conclusions are difficult and the AV often needs to be evaluated for individual experiment configurations, especially since the use of the RCMs can also degrade some aspects of the simulation, for example, amplifying biases already present in the driving GCMs [Laprise et al., 2008].

©2015. American Geophysical Union. All Rights Reserved.

TORMA ET AL.

Within the downscaling paradigm, RCMs are expected to add value mostly at scales not resolved by the driving GCMs [e.g., Giorgi et al., 2001; Laprise et al., 2008; Di Luca et al., 2012]; however, for large domains they can also improve the GCM information at large scales resolved by the driving GCM itself [e.g.; Giorgi et al., 1998; Veljovic et al., 2010; Diaconescu and Laprise, 2013]. In particular, regions for which RCM nesting is expected to provide substantial AV are areas characterized by complex topographical features varying at scales smaller than the resolution of GCMs. Within this context, RCMs are intended to produce improved representations of finescale spatial climate detail forced by topography. In addition, due to their enhanced spatial resolution, they are expected to provide AV in the simulation of the frequency distribution of weather events and extremes

ADDED VALUE OF RCM OVER COMPLEX TERRAIN

3957

Journal of Geophysical Research: Atmospheres

10.1002/2014JD022781

[Laprise et al., 2008; Giorgi, 2006]. Earlier work on the representation of topographical detail by RCMs has been indeed extensive (e.g., see references cited above), and in fact this was one of the main reasons for developing this modeling technique [e.g., Giorgi and Bates, 1989]. Recently, as part of a the international program CORDEX (Coordinated Regional Downscaling Experiment, http://wcrp-cordex.ipsl.jussieu.fr/) [Giorgi et al., 2009], and in particular its regional initiatives EUROCORDEX [Jacob et al., 2013] and MED-CORDEX (Ruti et al., MED-CORDEX initiative for Mediterranean climate studies, submitted to the Bulletin of the American Meteorological Society, 2014), ensembles of nested RCM experiments over the European and Mediterranean regions at two nominal horizontal resolutions, 0.44° (~50 km) and 0.11° (~12 km), have been completed. The availability of these data sets of unprecedented quality thus enables us to revisit the issue of AV in complex topography settings. In this study we focus on precipitation in the European Alps, a region not only common to the EURO-CORDEX and MED-CORDEX domains but also characterized by the availability of a high-quality, high-resolution gridded observed daily precipitation data set (the EURO4M-APGD) [Isotta et al., 2014]. Specifically, we analyze the simulation of precipitation in relation to three aspects for which the AV of the increased horizontal resolution of RCMs is expected to be particularly important over mountainous terrain: the spatial distribution of mean precipitation, the probability density function (PDF) of daily intensity events, and the fraction of precipitation accounted for by extremes. We first introduce quantitative metrics of the AV associated with each of these quantities and then calculate these metrics in ensembles of EUROCORDEX and MED-CORDEX RCM experiments, as well as in the driving GCM ones, for a historical simulated period (1976–2005, keeping in mind that since there is no assimilation of real data, i.e., the RCMs are driven by corresponding coupled GCM simulations, this does not represent the actual 1976–2005 period, and thus, the analysis can only be carried out in a climatological sense). Although the AV can be assessed for many and diverse variables, our focus on precipitation has three main motivations: (i) the strong potential of finding topographically forced AV, (ii) the availability of an observational data set of sufficient high quality to allow a reliable assessment of AV, and (iii) the importance of precipitation for impact assessment studies, which is one of the main reasons for using RCMs as downscaling tools. Our study adds to the current literature on AV in several respects. First, while the majority of previous work focused on reanalysis driven RCM experiments, we here focus on GCM-driven ones, thereby evaluating the RCM AV with respect to the driving GCMs. This is because most often the AV question is raised when the RCMs are driven by GCM fields and also due to the potential low quality of the driving GCM climatologies. Second, while previous work has mostly used single models and simulations, we here analyze ensembles of state-of-the-art GCM and RCM experiments, which should provide added robustness to our results. Finally, we use general quantitative metrics specifically aimed at identifying AV for variables and statistics where the AV is expected. We therefore hope that this set of metrics can provide a basic and intercomparable framework to assess AV in other regions. In addition, although the main focus of the paper is on the improvements obtained with RCMs in simulating present-day precipitation characteristics, we also present a brief analysis of whether the use of the RCMs also modulates the simulated precipitation change signal compared to that of the driving GCMs. We stress, however, that it is not the purpose of this paper to provide climate change scenarios for the region and that the limited discussion of the change patterns is presented only for illustrative purposes. In section 2 we first describe the observation and simulation data used in this work, along with the quantitative metrics calculated to measure the AV. We then discuss our results in section 3 and present our main conclusions in section 4.

2. Data and Methods 2.1. Observed and Model Data The observation data set we use for the assessment of the AV is the EURO4M-APGD described by Isotta et al. [2014]. It consists of daily precipitation station data regridded on a 5 km grid from 1971 to 2008 covering the domain shown in Figure 1. Specifically, we use data for the years 1976–2005, which represents our reference period. The gridded data were obtained by interpolation from about 8500 ground stations [Isotta et al., 2014],

TORMA ET AL.


3958


10.1002/2014JD022781

and the data set does not include any correction for wind-induced undercatch by precipitation gauges. Gauge undercatch, however, has been shown to be quite important, especially during the cold season and in mountainous areas, yielding an underestimate of up to 20–25% of real precipitation [Richter, 1995; Adam and Lettenmaier, 2003]. In order to account for this effect, in our analysis of mean precipitation, we first estimated a mean monthly undercatch correction from the global precipitation data set of the University of Delaware (UDel Version 3.01) [Legates and Willmott, 1990], by subtracting the uncorrected from the corrected data in their database at each UDel grid box. We then added this spatially varying correction at the monthly scale to the mean precipitation values in the EURO4M-APGD. We did not add this correction in the daily precipitation intensity and extreme analysis since only mean monthly climatological values are provided in the UDel data set. In terms of model simulations, we selected five sets of driving GCM and corresponding nested RCM experiments at 0.44° and 0.11° resolution (referred to as RCM44 and RCM11, respectively). These are reported in Table 1 and include three sets of runs from the EURO-CORDEX database and two from Figure 1. Topography for the Alpine analysis region at the three resolutions the MED-CORDEX one, with four driving investigated in this work: (top) 1.32°, (middle) 0.44°, and (bottom) 0.11°. GCMs from the CMIP5 experiment Units are in meters. (Climate Model Intercomparison Project, Phase 5) [Taylor et al., 2012] and five nested RCMs (note that the RCA4 and RegCM4 runs were driven by the same GCM). Each RCM experiment is carried out at both resolutions with the same model. More experiments were available at the time of completion of this work, particularly from EURO-CORDEX. However, they were performed with the same RCMs as in Table 1 (but with different forcing GCMs), and we decided to only select one run per RCM in order to prevent the results from a given RCM dominating the ensemble. Since the models (both the GCMs and RCMs) do not use the same horizontal grid, we interpolated all the original precipitation data onto three representative regular latitude/longitude grids having, respectively, a grid spacing of 0.11°, 0.44°, and 1.32°. The first two resolutions are close to those of the RCM runs, while the third resolution is representative of the mean resolution of the four driving GCM experiments. The observation data were also interpolated onto these common grids, allowing us a direct assessment and intercomparison of the results across resolutions; i.e., we assess not only the AV of the RCMs at the fine scale but also the (eventual) AV when the RCM fields are upscaled at the GCM-like resolution. The interpolation of data to the three common resolution grids was performed using the distance-weighted average remapping method in the Climate Data Operators software (CDO, https://code.zmaw.de/projects/cdo),

TORMA ET AL.


3959


10.1002/2014JD022781

a

Table 1. Overview of Global and Regional Climate Models Used in the Present Study Model a, CNRM-CM5

b, EC-EARTH c, HadGEM2-ES d, MPI-ESM-LR ALADIN (a-MC) CCLM (d-EC) RCA4 (c-EC) RACMO (b-EC) RegCM4 (c-MC)

Modeling Group

Resolution

Number of Vertical Levels

Reference

Centre National de Recherches Meteorologiques and Centre Europeen de Recherches et de Formation Avancee en Calcul Scientifique, France Irish Centre for High-End Computing, Ireland Met Office Hadley Centre, UK Max Planck Institute for Meteorology, Germany Centre National de Recherches Meteorologiques, France Climate Limited-area Modeling Community, Germany Swedish Meteorological and Hydrological Institute, Rossby Centre, Sweden Royal Netherlands Meteorological Institute, The Netherlands International Centre for Theoretical Physics, Italy

1.40625° × 1.40625°

31

Voldoire et al. [2012]

1.125° × 1.125° 1.875° × 1.2413° 1.875° × 1.875° 0.44°/0.11° 0.44°/0.11° 0.44°/0.11°

62 38 47 31 40 40

Hazeleger et al. [2010] Collins et al. [2011] Jungclaus et al. [2010] Colin et al. [2010] Rockel et al. [2008] Kupiainen et al. [2011]

0.44°/0.11° 0.44°/0.11°

40 23

Meijgaard et al. [2012] Giorgi et al. [2012]

a

For the regional models the letter in parenthesis indicates the driving GCM (from CMIP5) and whether the run uses the EURO-CORDEX (EC) or MED-CORDEX (MC) domain.

which utilizes the four nearest neighbor values for each grid cells from the original field. In order to minimize possible errors introduced in the upscaling and downscaling interpolations, the distance-weighted technique was applied gradually on the data, i.e., by using intermediate resolution grids between the 0.11°, 0.44°, and 1.32° grids (i.e., 0.22°, 0.66°, and 0.88°). We carefully compared several other available interpolation methods in the CDO software, such as bilinear, bicubic, and field conserving (REMAPCON), and found that the distance-weighted one produced the most consistent spatial patterns across resolutions. In addition, we found that on the monthly scale, our method conserved the area averaged total precipitation with an error of less than 1–2% between the 0.11° and 0.44° grids and less than 7–8% between the 0.11° and 1.32° grids, likely because of the use of the intermediate grids. These errors were similar to those found, for example, using the nominally mean field-conserving REMAPCON method. We thus concluded that our choice of interpolation procedure does not affect the main conclusions of our work, although we recognize that some uncertainty is present in our analysis due to the interpolation method selected, particularly when upscaling data from the fine 0.11° grid to the coarse 1.32° grid. Figure 1 shows our analysis area at the three common grid resolutions, which is the area covered by the EURO4M-APGD observation data, along with the corresponding representations of the topography of the Alps (obtained from an original data set at 30 s resolution, GTOPO30) [Bliss and Olsen, 1996]. Figure 1 thus illustrates the enhancement in the representation of the orographic spatial detail as the horizontal resolution increases, although it should be kept in mind that the RCM domains are characterized by different smoothings of the topographical field for numerical stability purposes, and thus, the RCMs use somewhat different topography representations even at the same nominal resolution. 2.2. Metrics of Added Value In our search for AV, we focus on the spatial distribution of mean precipitation, the daily precipitation PDFs, and the extremes tail of the daily PDF distribution, as measured by the fraction of precipitation accounted for by events above the 95th percentile, or R95 [e.g., Sillmann et al., 2013; Giorgi et al., 2014]. To portray the AV for the spatial precipitation distribution, we selected the Taylor diagram which, given two fields (in this case simulations and observations), provides information on multiple metrics: the spatial correlation, the spatial standard deviation, and the centered Root Mean Square Distance (C-RMSD, i.e., the RMSD with the bias removed). The Taylor diagram is especially illustrative of the spatial AV because it extracts multiple information specific to the spatial patterns, where indeed the AV of downscaling is expected. For the PDFs we use as metric the two-sample Kolmogorov-Smirnov (KS) distance [e.g., Chakravarti et al., 1967], i.e., the maximum distance between two cumulative distribution functions (CDFs), which essentially measures the dissimilitary between the given samples. The KS distance varies between 0, when the distributions overlap, and 1 when they have no overlap. It is calculated from the formula: dKS ðF; GÞ ¼ supt∈ℝ jF ðGÞ GðtÞj; where F and G are the two CDFs and sup represents the supremum function [Chakravarti et al., 1967].

TORMA ET AL.


3960


10.1002/2014JD022781

Finally, for the R95 we use as metrics the mean values, which are expected to be improved at higher downscaled resolutions, and the pattern correlation. Note that R95 is not a measure of absolute extremes, but it is relative to a given PDF. We selected it because it is included in the World Meteorological Organization list of recommended extreme indices [Peterson, 2005] and it is a commonly used index in analyses of extremes [Sillmann et al., 2013; Giorgi et al., 2014]. By better describing extremes and PDFs, downscaling is expected to also better reproduce the fraction of precipitation associated with the extremes tail of events (i.e., R95). Figure 2. Annual cycle of ensemble average mean monthly precipitation averaged over the entire analysis regions of Figure 1 and the reference period 1976–2005 for the different ensembles of GCM and RCM simulations (GCMs on the 1.32° grid, RCM44 on the 0.44° grid, RCM11 on the 0.11° grid) and the gauge-corrected and gauge-uncorrected observations (interpolated on the 0.11° grid). Units are mm/d.

Note that we intercompare all simulations at all three resolution grids; i.e., we assess the AV in RCM downscaling as well as upscaling of fine-scale information provided by the RCMs, the latter information being important in providing indications on whether the downscaling AV is based on a better representation of physical processes (e.g., topographic forcing) rather than a simple disaggregation of large scale data. In addition, we analyze the ensemble average of the models in order to assess the overall behavior of the ensemble. We also tested the results for individual models and found them to be qualitatively in line with the ensemble average, even though they obviously had quantitative differences in the various metrics.

3. Results Prior to addressing the issue of AV, we briefly examine how the models reproduce mean precipitation over the region of interest throughout the year. This is done in Figure 2, which presents the annual cycle of region-averaged monthly precipitation in the GCM, RCM44, and RCM11 ensembles, along with both the gauge-corrected and gauge-uncorrected observations on the 0.11° grid. It can be seen that the region experiences two precipitation maxima, one in May-June and one in October-November. The model ensembles, and in particular the RCMs, reproduce this seasonal cycle. In general, the amounts of simulated precipitation increase with resolution, i.e., going from the GCM to the RCM44 and RCM11 ensembles. RCM-produced precipitation is overestimated compared to the uncorrected data, but more in line with the gauge-corrected observations, except for an overestimate in the warmer months. The GCMs are close to the uncorrected observed values in winter (when they underestimate the gauge-corrected amounts) and underestimate precipitation in July-November. Note that only the RCM11 ensemble captures the observed maximum in June, while the GCM and RCM44 ensembles displace it to May. In addition, the GCMs do not capture the October maximum. Overall, despite the occurrence of biases in different months, Figure 2 shows that the models analyzed here exhibit a good performance in reproducing mean precipitation over the region, in line with, or better than, previous experiments [Jacob et al., 2007; Rauscher et al., 2010]. 3.1. Added Value for the Spatial Distribution of Mean Precipitation in the Historical Period 1976–2005 Figures 3 and 4 show the distribution of mean winter (December-January-February, or DJF) and summer (June-July-August, or JJA) precipitation for the ensemble average of the GCM and RCM (both the RCM11 and RCM44) simulations interpolated onto the three different common grids, along with the corresponding fields from the EURO4M-APGD observation data set, both with and without gauge correction. The increase in precipitation detail with increasing resolution, which is clearly tied to the topographical features of the Alpine chain (see Figure 1), is visually evident, as is the deterioration of this signal at the coarsest resolution. At the highest resolution, precipitation maxima are found across the

TORMA ET AL.


3961


10.1002/2014JD022781

Figure 3. Mean (1976–2005) winter precipitation (December-January-February, or DJF) interpolated on the three resolution grids of Figure 1 for the ensembles of driving GCMs, RCM44, and RCM11 (top three rows, respectively) simulations and the EURO4M-APGD observations without and with undercatch gauge correction (see text). Units are mm/d.

entire Alpine chain, and particularly the western Alps and Jura mountains in DJF and the central Alps in JJA. Other topographically induced maxima occur in the northern Apennines of Italy and over the coastal areas of Slovenia and Croatia. Also evident is the effect of the undercatch correction in enhancing cold season precipitation at high elevations. The RCM11 ensemble reproduces this distribution quite well in both seasons, capturing most maxima, along with some of the more prominent minima (e.g., over the Central-Eastern Alps at the Italy-Austria border or the region between the Jura and Alpine chains). Already at the 0.44° resolution, some of this detail is lost, with the RCM44 models producing only one area of precipitation maximum centered over the western Alps in DJF and the central Alps in JJA. At the coarsest resolution, both the driving GCMs and the upscaled RCMs and observations show a very smooth precipitation gradient from the western to the eastern Alps in DJF and from the central Alps outward in JJA. Similar results were found for the other two seasons (not shown for brevity). Figures 3 and 4 thus illustrate the realistic fine-scale AV that can be obtained by the use of RCMs, as well as the gain achieved in increasing the RCM resolution. For a quantitative estimate of the AV, Figure 5 presents Taylor diagrams for the ensemble average precipitation in the four seasons over the area of interest at all three resolution grids. In the figure the red color refers to the RCM11 ensemble, the blue to the RCM44 ensemble, and the black to the driving GCM one, while the different shapes of the symbols indicate the different analysis grids. The Taylor diagram enables a comparison of the data at different resolutions, so that, for example, the black square indicates that the GCM fields interpolated onto the 0.44° grid are compared to the 0.44° observations, while the black circle compares the GCM fields onto the coarse 1.32° grid with the observations upscaled at that grid. As another example, the red triangle compares the

TORMA ET AL.


3962


10.1002/2014JD022781

Figure 4. Same as Figure 3, but for summer (June-July-August, or JJA).

high-resolution RCM11 and observations on the fine-scale 0.11° grid, while the red circle compares RCM11 and observations both upscaled at the 1.32° grid. In the analysis of Figure 5, we also note that the correlations at different resolutions are calculated using very different numbers of grid points (or degrees of freedom) and this affects the robustness of the correlation itself, this robustness being higher when more grid points are used (finer scale grid). Therefore, the intercomparison of the correlation coefficients across the different resolution grids can only be done in a qualitative way. Finally, in Figure 5 we use as reference the gauge-corrected observations, since this is likely more representative of the mean climatology over this mountainous region. When using the uncorrected data (not shown for brevity) the correlations are essentially the same, while the normalized standard deviations tend to increase in the cold seasons, but the conclusions are qualitatively the same as with the corrected data. With these caveats in mind, Figure 5 first shows that at all resolution grids the RCM11 (red symbols) exhibits the smallest C-RMSD (distance from the point 1); i.e., the AV of the highest resolution RCMs is found not only at fine scales but also when upscaled at GCM-like resolution. Specifically, it can be seen that as the resolution of the models increases (from GCMs to RCM44 to RCM11), the correlations increase, and this occurs at all three grids. If we take as an example DJF, the correlations are in the range of 0.3–0.5 for the GCMs, 0.55–0.67 for the RCM44, and 0.72–0.76 for the RCM11. Similar types of improvements are found for the other seasons. Intercomparison across resolutions of the results for each model ensemble indicates that for the GCMs, as expected, the best performance occurs at the coarsest scale (black circle), particularly because of the (expected) underestimate of spatial standard deviation at fine scales. For the RCMs, the Taylor diagram metrics show more comparable values, and thus performance, across resolutions, but in most cases the RCM44 scores are intermediate between the GCM and RCM11 ones.

TORMA ET AL.


3963


10.1002/2014JD022781

Figure 5. Taylor diagram of mean (1976–2005) seasonal precipitation for the three model ensembles (black color for the driving GCMs, blue color for the RCM44, and red color for the RCM11) versus the gauge-corrected EURO4M-APGD observations at the three resolution grids of Figure 1 (circles for the 1.32° grid, squares for the 0.44° grid, and triangles for the 0.11° grid). The four panels refer to the four seasons (December-January-February, or DJF; March-April-May or MAM; June-July-August, or JJA; and September-October-November, or SON).

In summary, the analysis of mean precipitation spatial pattern measured in terms of Taylor diagrams, and in particular pattern correlations, shows that the RCMs improve the driving GCM results, i.e., provide AV not only at fine scales (downscaling) but also when upscaled at the scale of the GCM resolution. Intercomparison of Figures 1, 3, and 4 then shows that this AV is clearly related to a better representation of the topographical forcing on precipitation. 3.2. Added Value for the Daily Precipitation PDFs in the Historical Period 1976–2005 In order to intercompare the PDFs of daily precipitation intensities, we use the daily data from each simulation interpolated onto the three common grids of Figure 1 and include together all events of the historical time series (i.e., do not separate the events by season). Figure 6 presents the PDFs of all individual simulations at each resolution grid, along with the corresponding observed values. The PDFs for the fine-scale 0.11° grid clearly illustrate the AV of the RCM-based downscaling (RCMs versus GCMs) and of the increased RCM resolution (RCM11 versus RCM44). The GCMs only simulate events of up to about 90 mm/d (Figure 6c), while the observations interpolated at the 0.11° grid show events reaching 340 mm/d (Figure 6a). Conversely, the GCMs overestimate the frequency of events of very low intensities. The RCM44 ensemble already substantially improves the PDFs compared to the GCMs, but a further improvement is found for the RCM11 ensemble. On average, the RCM11 models still appear to underestimate the high-intensity tail of the observed distribution and overestimate the low intensity one, but the range of RCM11 PDFs almost encompasses the observed one, and in particular the RCM11 models do produce events of the highest magnitudes (exceeding 300 mm/d) as also found in the observations. A similar downscaling AV of GCM versus RCM daily precipitation PDFs at fine scales was found by Giorgi et al. [2014] for simulations over different CORDEX domains.

TORMA ET AL.


3964


10.1002/2014JD022781

0

10

GCM − 0.11 RCM44 − 0.11 RCM11 − 0.11 EURO4M−APGD − 0.11

−2

Frequency

10

−4

10

−6

10

−8

10 10

−10

0

50

100

150

200

250

300

350

400

0

10


−2

Frequency

10

−4

10

−6

10

−8

10 0

50

100

150

200

250

300

350

400

0

Frequency

10


−2

10

−4

10

−6

10 0

10

20

30

40

50

60

70

80

90

100

Intensity [mm/day] Figure 6. Daily precipitation intensity empirical probability distribution functions (PDFs) (frequency versus intensity of daily precipitation events, 1976–2005) for all the model experiments (black color for the driving GCMs, blue color for the RCM44, and red color for the RCM11) and the EURO4M-APGD observations (green color) interpolated onto the three resolution grids of Figure 1: (a) 0.11° grid, (b) 0.44° grid, and (c) 1.32° grid.

An interesting aspect of Figure 6 is that an analysis of the PDFs upscaled at the coarser resolution grids also reveals significant AV by the RCMs. In particular, even at the coarsest 1.32° grid, the GCMs appear to underestimate the occurrence of medium- to high-intensity events and overestimate the occurrence of light events, while both the RCM44 and RCM11 ensembles are closer to observations. By the same token, the RCM11 PDFs are generally closer to observations than the RCM44 ones even at the 0.44° resolution grid. In other words, also for this metric the highest-resolution RCMs (0.11°) add value not only at the finest scale but also when upscaled at the coarse scale of the GCMs or the intermediate scale of the RCM44. For a more quantitative assessment of this conclusion, Figure 7 presents the ensemble mean KS distance between simulated and observed CDFs for all GCM and RCM ensembles at the three resolution grids. The ensemble mean was obtained by first calculating the KS distance for each model run and then averaging over all models in the ensemble. The AV obtained with the RCM11 simulations is evident in that at all three resolution grids, the KS distance is lowest for the RCM11 ensemble. We also find that this distance TORMA ET AL.


3965


Kolmogorov-Smirnov distance

0.5

10.1002/2014JD022781

increases for each ensemble as the resolution increases, which however is likely due to the fact that tail of the distribution considerably expands with resolution. Finally, as with the previous metric, the RCM44 results are generally intermediate between the GCM and RCM11 ones.

0.4

0.3

0.2

0.1

0 1.32

0.44

GCM

RCM44

0.11

RCM11

Figure 7. Ensemble mean Kolmogorov-Smirnov (KS) distance (see text) for the driving GCM (black), RCM44 (blue), and RCM11 (red) daily precipitation CDFs versus the observed CDFs for the 1976–2005 historical period on the three resolution grids of Figure 1.

Summarizing the results of this section, we again find a substantial AV by the RCMs compared to the driving GCMs in the simulation of daily intensity PDFs, along with an increasing performance at higher resolutions in the RCM ensembles. This AV is evident not only at the downscaled fine scales but also when upscaled at the coarsest scales.

3.3. Added Value for Extreme Precipitation Events Tail (R95) in the Historical Period 1976–2005 The third metric we consider is the extreme precipitation tail as measured by the R95 defined in section 2. As for the PDFs, here the R95 metric is calculated on a full annual basis, i.e., including all days in the historical period simulations. Also, in the calculation of the R95, only days with at least 1 mm/d are included. Figure 8 shows the spatial distribution of R95 on the three resolution grids for the models’ ensemble average, while Figure 9 shows the corresponding spatial correlation coefficients between simulated and observed R95 at the different resolutions.

Figure 8. Ensemble mean daily precipitation R95 for the model ensembles (driving GCMs, RCM44, and RCM11 in the top three rows, respectively) and the EURO4M-APGD observations (bottom row) for the historical period 1976–2005 and the three resolution grids of Figure 1. Units are in percent of total precipitation accounted for by events above the 95th percentile.

TORMA ET AL.


3966


10.1002/2014JD022781

1

Correlation

The observed value of R95 at the 0.11° grid varies mostly in the range of 18– 0.8 30%, with peaks exceeding 30% in the southern flanks of the western Alps 0.6 (border between Switzerland and Italy) and a small region in northeastern 0.4 Italy (Friuli region). Other areas of high R95 include the coastal regions 0.2 of the Gulf of Genoa, which is one of the main cyclogenesys regions in the Mediterranean, and parts of 0 1.32 0.44 0.11 northeastern Italy. North of the Alps, GCM RCM44 RCM11 the R95 has generally lower values. By and large, the RCM11 ensemble Figure 9. Spatial correlation coefficient between simulated (ensemble captures well the magnitude and average) and observed daily precipitation R95 for the driving GCM (black), locations of these R95 maxima and RCM44 (blue), and RCM11 (red), and the EURO4M-APGD observations for minima, while the driving GCM the historical period 1976–2005 and the three resolution grids of Figure 1. ensemble entirely misses this highresolution detail, showing only a southwestward gradient in R95 and exhibiting generally lower values of R95 than observed at fine scales. The region-averaged R95 at the respective resolutions are 19.1% for the GCMs, 21.5% for the RCM44, 22.1% for the RCM11, and 22.2% for the EURO4M-APGD, again indicating a better performance for the RCM11 ensemble. An examination of the pattern correlations of Figure 9 clearly reveals the AV of the RCM downscaling. Both at the 0.44° and 0.11° grids, the correlation coefficients increase from the GCMs (about 0.5) to the RCM44 and RCM11, for which they are about 0.8. At the coarsest scale, all the ensembles show large and comparable correlation values of about 0.9. We also note that the correlations tend to decrease with increasing resolution of the grid, probably a result of the greater number of degrees of freedom (grid points) at the finer scales. In summary, for the R95 metric the RCM11 ensemble shows the best performance at the 0.11° and 0.44° resolutions compared to either the driving GCMs or the RCM44 ensembles, an indication of AV by the RCM downscaling and high model resolution. For this metric the results are comparable across the three ensembles when upscaled at the coarsest resolution grid. In addition, we carried the same R95 analysis for the four seasons separately and found results in line with those obtained for the annual R95 (see supporting information Table S1 and Figures S1–S8). 3.4. Changes in Mean Precipitation and R95 (2070–2099 Versus 1976–2005) Having assessed the AV of the three metrics selected for the historical period 1976–2005, we now turn our attention to the simulated changes for a future time period, 2070–2099, with respect to the historical one. The purpose of this analysis is not to provide a climate change scenario for the region but more simply to assess whether the high resolution of the nested RCMs affects the climate change signal compared to that of the driving GCMs through some illustrative examples related to the AV metrics selected. Toward this purpose we analyzed for the GCM, RCM44, and RCM11 ensembles, the 2070–2099 time slice of 21st century projections under the high-end RCP8.5 greenhouse gas concentration scenario [Moss et al., 2010]. Here we limit our analysis to the mean precipitation and R95, since the assessment of changes in PDFs requires a more detailed approach. Figures 10 and 11 first show the change in ensemble mean precipitation over the Alpine domain of interest at all three resolution grids. In general, the large-scale precipitation change follows the well-known pattern described for example by Giorgi and Coppola [2009], consisting of a pattern of increased precipitation over northern Europe and decreased precipitation over southern Europe. This pattern moves in a northsouth direction throughout the seasonal cycle following the corresponding motion of an anticyclonic sea level pressure increase. As a result of this change pattern, most of the Alpine region exhibits a positive precipitation change in the winter and a negative change in the summer, with a north-south gradient (from positive to the north to negative in the south) in the intermediate seasons.

TORMA ET AL.


3967


10.1002/2014JD022781

Figure 10. Ensemble mean precipitation change (2070–2099 minus 1976–2005) for the driving (top) GCM, (middle) RCM44, and (bottom) RCM11 simulations and for the three resolution grids of Figure 1. (a) DJF (winter) and (b) MAM (Spring). Units are in percent of precipitation amount during the historical period 1976–2005.

At the coarse 1.32° scale, the change pattern described above is essentially followed by all three ensembles, however with differences inherited by the upscaling from the fine-resolution runs. As we move to the highresolution grids, differences between the driving GCM and RCM ensembles emerge more strongly. For example, in DJF the RCMs show a tongue of reduced precipitation change extending from the Gulf of Genoa northeastward across the northwestern flanks of the Alpine chain. In March-April-May (MAM) an area of reduced precipitation appears in the RCM ensembles extending across the coastal regions of the Gulf of Genoa. Greater differences are then found in JJA and September-October-November (SON). In summer, over the central Alpine mountains, the large-scale reduction in precipitation found in the driving GCMs is weakened in the RCM44 ensemble and turns even positive (increase in precipitation) in the RCM11 one. This result, found also in previous high-resolution RCM scenarios [Gao et al., 2006; Im et al., 2010; Fischer et al., 2014], has been tied to increased convective activity at the highest mountainous elevations under warming conditions [Fischer et al., 2014]. In SON, the RCM11 ensemble projects a marked decrease in precipitation throughout most of the Alpine chain, while the GCM ensemble shows a precipitation increase or a small change. Such differences have been tied to changes in moisture flux with respect to the orientation of major mountain chains [Gao et al., 2006]; however, a detailed analysis of them

TORMA ET AL.


3968


10.1002/2014JD022781

Figure 11. Same as Figure 10 but for (a) JJA (summer) and (b) SON (autumn).

is beyond the scope of the present paper. It is then clear from Figures 10 and 11 that the RCM-simulated differences in precipitation change (compared to the driving GCMs) are also reflected in the upscaled results at the GCM resolution. Figure 12 shows modified versions of the Taylor diagram aimed at measuring differences between the spatial patterns of the changes in the driving GCM and nested RCM ensembles. In these diagrams the role of “observations” is played by the ensemble average GCM change, and the ensemble average RCM changes (RCM44 and RCM11) are compared to the corresponding driving GCM ones at each of the three resolution grids. It is evident that large differences are found between the ensemble average GCM and RCM changes, which generally increase from the RCM44 to the RCM11 ensembles. In fact, these differences are evident not only at the finest resolution grid but also when the RCM fields are upscaled at the GCM scale. Overall, the higher-resolution RCMs mostly produce larger spatial variability of change than the driving GCMs (standard deviation ratio greater than 1). The correlation between the RCM11 and GCM changes are as low as 0.5 in DJF and 0.2 in SON, with intermediate values for the RCM44 runs. Figure 13 shows the ensemble average R95 changes in the GCM, RCM44, and RCM11 ensembles at the different resolution grids. Although all the ensembles project a predominant increase of R95, a consistent signal found in most GCM and RCM simulations [e.g., Sillmann et al., 2013; Giorgi et al., 2014], differences in

TORMA ET AL.


3969


10.1002/2014JD022781

Figure 12. Taylor diagram of the ensemble mean precipitation change (2070–2099 minus 1976–2005) comparing the RCM44 (blue) and RCM11 (red) changes with the corresponding driving GCM ensemble average change for the three resolution grids of Figure 1.

the patterns of change are marked, particularly in correspondence of the Alpine chain. For example, the RCM11 ensemble projects small pockets of decrease in R95 over the northern flanks of the western Alps, where the driving GCMs still project an increase. Similarly, over the Piemonte regions south of the western Alpine chain, the RCMs locate an area of maximum R95 increase not found in the driving GCMs. Indeed,

Figure 13. Ensemble mean change in R95 (2070–2099 minus 1976–2005) for the driving (top) GCM, (middle) RCM44, and (bottom) RCM11 experiments and the three resolution grids of Figure 1. Units are in percent of total precipitation accounted for by daily events of intensity above the 95th percentile.

TORMA ET AL.


3970


10.1002/2014JD022781

the correlations between the GCM and RCM ensemble mean R95 changes vary from ~0.7 at the 1.32° resolution grid to 0.4–0.5 at the finer resolution ones. When averaged over the entire region, the changes in R95 are also different across the ensembles, having magnitudes of 1.37% for the GCMs, 1.31% for the RCM44, and 1.07% for the RCM11. We do not analyze in detail here the specific reasons for these differences in change, which requires a separate detailed work, but the results of this section show that the RCM downscaling produces a substantial modulation of the GCM precipitation change signal at various scales over the Alpine region.

4. Summary and Final Considerations In this paper we investigated the AV of dynamical downscaling by RCMs compared to the driving GCMs over the mountainous terrain of the European Alps. We assessed different AV metrics related to the spatial patterns of mean precipitation, daily precipitation PDFs, and precipitation extremes tail (R95) in two ensembles of RCM simulations at the nominal resolutions of 0.11° and 0.44° produced as part of the EUROCORDEX and MED-CORDEX programs. Our region of interest is the Alpine area, for which a high-quality, fine-resolution gridded observation data set is available (the EURO4M-APGD described by Isotta et al. [2014]), and we intercompared driving GCM and nested RCM ensembles at three different resolution grids —0.11°, 0.44°, and 1.32°. This enables us to assess not only the AV of downscaling but also that of upscaling the RCM results to the GCM scale. We found substantial AV of RCM downscaling in all precipitation metrics considered, with this AV increasing with the resolution of the RCMs (i.e., being maximum for the RCM11 ensemble). The AV was clearly associated to the fine-scale topography characterizing this region. In addition, the RCM results were mostly better than the driving GCM ones for the metrics considered also when upscaled at the coarse GCM scale. This adds robustness to the conclusion that the AV obtained with the RCM downscaling is due to physically consistent processes (i.e., topographical forcing on precipitation) and not simply to a high-resolution disaggregation of the GCM fields. We also showed that the fine-scale topographical detail of the Alpine chain substantially modulates the precipitation change signal in the RCMs compared to those in the driving GCMs, sometimes even in the sign of the change.

Acknowledgments We thank the CMIP5, EURO-CORDEX, and MED-CORDEX modeling groups for making available the simulation data used in this work and the Swiss Federal Office for Meteorology and Climatology along with the National Oceanic and Atmospheric Administration (NOAA) for providing the EURO4M-APGD and UDel observation data sets. The data used in this work can be found at the following web sites: http://cordexesg.dmi.dk/ esgf-web-fe/ (EURO-CORDEX), http:// www.medcordex.eu/medcordex.php (MED_CORDEX), http://cmip-pcmdi.llnl. gov/cmip5/data_portal.html (CMIP5), http://www.euro4m.eu/datasets.html (EURO4M-APGD), and http://www.esrl. noaa.gov/psd/data/gridded/data. UDel_AirT_Precip.html (UDel).

TORMA ET AL.

Our results provide a clear and strong evidence of the AV deriving from the use of high-resolution RCMs in areas of complex mountainous terrain. The AV metrics we employed are of general purpose and specifically aimed at model aspects for which the AV is expected to be important. Therefore, they could be used to analyze the issue of AV over other regions of the globe, in particular when complex topography is a strong modulating factor of the surface climate signal (e.g., the western U.S.). The fact that the precipitation change signal in the high-resolution RCMs shows patterns substantially different from those of the driving GCMs points to the caution that needs to be taken when using GCM projections for regional impact applications. In this regard, RCM-derived projections can add information of strong value. Our work thus highlights the important role that high-resolution RCMs can play in the study of climate change over regions characterized by complex topography, particularly concerning fine-scale spatial information and higher-order climate statistics, which are important for a multiplicity of impacts sectors, such as related to water resources, agriculture, and natural disaster prevention.

References Adam, J. C., and D. P. Lettenmaier (2003), Adjustment of global gridded precipitation for systematic bias, J. Geophys. Res., 108(D9), 4257, doi:10.1029/2002JD002499. Bliss, N. B., and L. M. Olsen (1996), Development of a 30-arc-second digital elevation model, in Pecora Thirteen, Human Interactions With the Environment—Perspectives from Space, Sioux Falls, South Dakota, 20–22 August. Castro, C. L., R. A. Pielke, and G. Leoncini (2005), Dynamical downscaling: An assessment of value added using a regional climate model, J. Geophys. Res., 110, D05108, doi:10.1029/2004JD004721. Chakravarti, I. M., R. G. Laha, and J. Roy (1967), Handbook of Methods of Applied Statistics, Volume I, pp. 392–394, John Wiley, Hoboken, N. J. Colin, J., M. Déqué, R. Radu, and S. Somot (2010), Sensitivity study of heavy precipitations in Limited Area Model climate simulation: Influence of the size of the domain and the use of the spectral nudging technique, Tellus A, 62, 591–604. Collins, W. J., et al. (2011), Development and evaluation of an Earth-System model—HadGEM2, Geosci. Model Devel., 4, 1051–1075. Coppola, E., F. Giorgi, S. Rauscher, and C. Piani (2010), Development of regional climate model weights based on the model’s “mesoscale signal”, Climate Res., 44, 121–134. De Sales, F., and Y. Xue (2010), Assessing the dynamic downscaling ability over South America using the intensity-scale verification technique, Int. J. Climatol., 31, 1254–1263.


3971


10.1002/2014JD022781

Di Luca, A., R. de Elia, and R. Laprise (2012), Potential for added value in precipitation simulated by high resolution nested Regional Climate Models and observations, Clim. Dynam., 38, 1229–1247. Di Luca, A., R. de Elia, and R. Laprise (2013), Potential for small scale added value of RCM’s downscaled climate change signal, Clim. Dyn., 40, 1415–1433. Diaconescu, E. P., and R. Laprise (2013), Can added value be expected in RCM-simulated large scales?, Clim. Dyn., 41, 1769–1800. Feser, F. (2006), Enhanced detectability of added value in limited-area model results separated into different spatial scales, Mon. Weather Rev., 134, 2180–2197. Fischer, A. M., et al. (2014), Projected changes in precipitation intensity and frequency in Switzerland: A multi-model perspective, Int. J. Climatol., doi:10.1002/joc.4162. Gao, X. J., J. S. Pal and F. Giorgi (2006), Projected changes in mean and extreme precipitation over the Mediterranean region from high resolution double nested RCM simulations, Geophys. Res. Lett., 33, L03706, doi:10.1029/2005GL024954. Giorgi, F. (2006), Regional climate modeling: Status and perspectives, J. Phys. IV, 139, 101–118. Giorgi, F., and E. Coppola (2009), Projections of 21st century climate over Europe, European Physical Journal, Web of Conferences, 1, pp. 29–46. Giorgi, F., and G. T. Bates (1989), The climatological skill of a regional model over complex terrain, Mon. Weather Rev., 117, 2325–2347. Giorgi, F., C. Shields-Brodeur, and G. T. Bates (1994), Regional climate change scenarios over the United States produced with a nested regional climate model, J. Clim., 7, 375–399. Giorgi, F., L. O. Mearns, C. Shields, and L. McDaniel (1998), Regional nested model simulations of present day and 2XCO2 climate over the Central Plains of the U.S, Clim. Change, 40, 457–493. Giorgi, F., et al. (2001), Regional climate information—Evaluation and projections, in Climate Change 2001: the Scientific Basis. Contribution of Working Group I to the Third Assessment Report of the Intergovernmental Panel on Climate Change, edited by J. T. Houghton, pp. 583–638, Cambridge Univ. Press, Cambridge, U. K., and New York. Giorgi, F., C. Jones, and G. Asrar (2009), Addressing climate information needs at the regional level: The CORDEX framework, WMO Bull., 58, 175–183. Giorgi, F., et al. (2012), RegCM4: Model description and preliminary tests over multiple CORDEX domains, Clim. Res., 52, 7–29. Giorgi, F., et al. (2014), Changes in extremes and hydroclimatic regimes in the CREMA ensemble projections, Clim. Change, 125, 39–51. Hazeleger, W., et al. (2010), EC-Earth: A seamless Earth-System prediction approach in action, Bull. Am. Meteorol. Soc., 91, 1357–1363. Im, E.-S., E. Coppola, F. Giorgi, and X. Bi (2010), Local effects of climate change over the Alpine region: A study with a high resolution regional climate model with a surrogate climate change scenario, Geophys. Res. Lett., 37, L05704, doi:10.1029/2009GL041801. Isotta, F. A., et al. (2014), The climate of daily precipitation in the Alps: Development and analysis of a high-resolution grid dataset from pan-Alpine rain gauge data, Int. J. Climatol., 34, 1657–1675. Jacob, D., et al. (2007), An inter-comparison of regional climate models for Europe: Ddesign of the experiments and model performance, Clim. Change, 81, 31–52. Jacob, D., et al. (2013), EURO-CORDEX: New high resolution climate change projections for European impact research, Reg. Environ. Change, 14, 563–578. Jungclaus, J. H., et al. (2010), Climate and carbon-cycle variability over the last millennium, Clim. Past, 6, 723–737. Kanamitsu, M., and H. Kanamaru (2007), Fifty-seven-year California reanalysis downscaling at 10 km (CaRD10). Part I: System detail and validation with observations, J. Clim., 20, 5553–5571. Kerkhoff, C., H. R. Kuensch, and C. Schaer (2014), Assessment of bias assumptions for climate models, J. Clim., 27, 6799–6818. Kupiainen, M., et al. (2011), Rossby Centre Regional Atmospheric Model, RCA4, Rossby Centre Newsletter, 54 pp. Laprise, R. (2014), Comment on “The added value to global model projections of climate change by dynamical downscaling: A case study over the continental U.S. using the GISS-ModelE2 and WRF models” by Racherla et al, J. Geophys. Res. Atmos., 119, 3877–3881, doi:10.1002/ 2013JD019945. Laprise, R., et al. (2008), Challenging some tenets of Regional Climate Modelling, Meteorol. Atmos. Phys., 100, 3–22. Legates, D. R., and C. J. Willmott (1990), Mean seasonal and spatial variability in gauge-corrected, global precipitation, Int. J. Climatol., 10, 111–127. Leung, L. R., L. O. Mearns, F. Giorgi, and R. L. Wilby (2003), Regional climate research: Needs and opportunities, Bull. Am. Meteorol. Soc., 82, 89–95. Mariotti, L., E. Coppola, M. B. Sylla, F. Giorgi, and C. Piani (2011), Regional climate model simulations of projected 21st century climate change over an all-Africa domain: Comparison analysis of nested and driving model results, J. Geophys. Res., 116, D15111, doi:10.1029/2010JD015068. Mariotti, L., I. Diallo, E. Coppola, and F. Giorgi (2014), Seasonal and intraseasonal changes of Africa monsoon climates in 21st century CORDEX projections, Clim. Change, 125, 53–65. Meijgaard E., et al. (2012), Refinement and application of a regional atmospheric model for climate scenario calculations of Western Europe, Climate Changes Spatial Planning Publication: KvR 054/12, pp 44. Moss, R. H., et al. (2010), The next generation of scenarios for climate change research and assessment, Nature, 463(7282), 747–756. Peterson, T. C. (2005), Climate change indices, WMO Bull., 54, 82–86. Prömmel, K., B. Geyer, J. M. Jones, and M. Widmann (2010), Evaluation of the skill and added value of a reanalysis-driven regional simulation for Alpine temperature, Int. J. Climatol., 30, 760–773. Racherla, P. N., D. T. Shindell, and G. S. Faluvegi (2012), The added value to global model projections of climate change by dynamical downscaling: A case study over the continental U.S. using the GISS-ModelE2 and WRF models, J. Geophys. Res., 117, D20118, doi:10.1029/2012JD018091. Rauscher, S. A., E. Coppola, C. Piani, and F. Giorgi (2010), Resolution effect of regional climate model simulation of precipitation over Europe. Part I: Seasonal, Clim. Dyn., 35, 685–711. Richter D. (1995), Ergebnisse methodischer Untersuchungen zur Korrektur des systematischen Messfehlers des Hellmann-Niederschlagsmessers, Bericht Deutschen Wetterdienstes, 194, 93 pp. Rockel, B., A. Will, and A. Hense (2008), Special issue regional climate modelling with COSMO-CLM (CCLM), Meteorol. Z., 17, 347–348. Sillmann, J., et al. (2013), Climate extreme indices in the CMIP5 multi-model ensemble: Part 2. Future climate projections, J. Geophys. Res. Atmos., 118, 2473–2493, doi:10.1002/jgrd.50188. Sotillo, M. G., R. Aznar, and F. Valero (2005), The 44-year Mediterranean HIPOCAS wind database: A useful tool to analyse offshore extreme wind events from a long-term regional perspective, Coast. Eng., 55, 930–943. Taylor, K. E., R. J. Stouffer, and G. A. Meehl (2012), An overview of CMIP5 and the experiment design, Bull. Am. Meteorol. Soc., 93, 485–498. Veljovic, K., B. Rajkovic, M. J. Fennessy, E. L. Altshuler, and F. Mesinger (2010), Regional climate modeling: Should one attempt improving on the large scales? Lateral boundary condition scheme: Any impact?, Meteorol. Z., 19, 237–246. Voldoire, A., et al. (2012), The CNRM-CM5 global climate model: Description and basic evaluation, Clim. Dyn., 40, 2091–2121. Winterfeldt, J., and R. Weisse (2009), Assessment of value added for surface marine wind speed obtained from two regional climate models, Mon. Weather Rev., 137, 2955–2965.

TORMA ET AL.


3972